On Fri, Jan 28, 2022 at 09:26:09AM +0100, Magnus Groß wrote:
> On Wed, Jan 26, 2022 at 10:31:42PM -0800 Kees Cook wrote:
> > On Wed, Jan 26, 2022 at 08:50:15AM -0800, Kees Cook wrote:
> > > On Wed, Jan 26, 2022 at 05:25:20PM +0100, Magnus Groß wrote:
> > > > From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
> > > > From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <magnus.gross@xxxxxxxxxxxxxx>
> > > > Date: Wed, 26 Jan 2022 16:35:07 +0100
> > > > Subject: [PATCH] elf: Relax assumptions about vaddr ordering
> > > > MIME-Version: 1.0
> > > > Content-Type: text/plain; charset=UTF-8
> > > > Content-Transfer-Encoding: 8bit
> > > >
> > > > Commit 5f501d555653 ("binfmt_elf: reintroduce using
> > > > MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now
> > > > assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().
> > > >
> > > > Specifically consider an ELF binary with the following PT_LOAD segments:
> > > >
> > > > Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
> > > > LOAD 0x000000 0x08000000 0x08000000 0x474585 0x474585 R E 0x1000
> > > > LOAD 0x475000 0x08475000 0x08475000 0x090a4 0xc6c10 RW 0x1000
> > > > LOAD 0x47f000 0x00010000 0x00010000 0x00000 0x7ff0000 0x1000
> > > >
> > > > Note how the last segment is actually the first segment and vice versa.
> > > >
> > > > Since total_mapping_size() only computes the difference between the
> > > > first and the last segment in the order that they appear, it will return
> > > > a size of 0 in this case, thus causing load_elf_binary() to fail, which
> > > > did not happen before that change.
> > > >
> > > > Strictly speaking total_mapping_size() made that assumption already
> > > > before that patch, but the issue did not appear because the old
> > > > load_addr_set guards never allowed this call to total_mapping_size().
> > > >
> > > > Instead of fixing this by reverting to the old load_addr_set logic, we
> > > > fix this by comparing the correct first and last segments in
> > > > total_mapping_size().
> > >
> > > Ah, nice. Yeah, this is good.
> > >
> > > > Signed-off-by: Magnus Groß <magnus.gross@xxxxxxxxxxxxxx>
> > >
> > > Fixes: 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > Acked-by: Kees Cook <keescook@xxxxxxxxxxxx>
> >
> > Andrew, can you pick this up too?
> >
> > -Kees
> >
>
> May I also propose to include this patch in whatever mailing-list
> corresponds to the 5.16.x bugfix series?
> It turns out that almost all native Linux games published by the Virtual
> Programming company have this kind of weird PT_LOAD ordering including
> the famous Bioshock Infinite, so right now those games are all
> completely broken since Linux 5.16.
Thanks for additional rationale!
I included the field that would be expected to have this picked up for
the 5.16 stable tree (the "Fixes: ..." and "Cc: stable@xxxxxxxxxxxxxxx")
so once it lands in Linus's tree, it'll get picked up for the v5.16.z
series too.
> P.S.: Someone should probably ask Virtual Programming, what kind of
> tooling they use to create such convoluted ELF binaries.
Does "strings" provide any hints? :)
--
Kees Cook
